import cv2 import torch import torchvision from osdsynth.processor.wrappers.grounding_dino import get_grounding_dino_model from osdsynth.processor.wrappers.ram import get_tagging_model, run_tagging_model from osdsynth.processor.wrappers.sam import ( convert_detections_to_dict, convert_detections_to_list, crop_detections_with_xyxy, filter_detections, get_sam_predictor, get_sam_segmentation_from_xyxy, mask_subtract_contained, post_process_mask, sort_detections_by_area, ) from osdsynth.utils.logger import SkipImageException from osdsynth.visualizer.som import draw_som_on_image from PIL import Image import numpy as np class SegmentImage: """Class to segment the image.""" def __init__(self, cfg, logger, device, init_gdino=True, init_tagging=True, init_sam=True): self.cfg = cfg self.logger = logger self.device = device if init_gdino: # Initialize the Grounding Dino Model self.grounding_dino_model = get_grounding_dino_model(cfg, device) else: self.grounding_dino_model = None if init_tagging: # Initialize the tagging Model self.tagging_transform, self.tagging_model = get_tagging_model(cfg, device) else: self.tagging_transform = self.tagging_model = None if init_sam: # Initialize the SAM Model self.sam_predictor = get_sam_predictor(cfg.sam_variant, device) else: self.sam_predictor = None pass def process(self, image_bgr, two_class ,plot_som=True): """Segment the image.""" image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB) image_rgb_pil = Image.fromarray(image_rgb) # image_rgb_pil.save('tmp.png') img_tagging = image_rgb_pil.resize((384, 384)) img_tagging = self.tagging_transform(img_tagging).unsqueeze(0).to(self.device) # Tag2Text if two_class is None: classes = run_tagging_model(self.cfg, img_tagging, self.tagging_model) else: classes = two_class if len(classes) == 0: raise SkipImageException("No foreground objects detected by tagging model.") # Using GroundingDINO to detect and SAM to segment detections = self.grounding_dino_model.predict_with_classes( image=image_bgr, # This function expects a BGR image... classes=classes, box_threshold=self.cfg.box_threshold, text_threshold=self.cfg.text_threshold, ) if len(detections.class_id) < 1: raise SkipImageException("No object detected.") # Non-maximum suppression nms_idx = ( torchvision.ops.nms( torch.from_numpy(detections.xyxy), torch.from_numpy(detections.confidence), self.cfg.nms_threshold, ) .numpy() .tolist() ) print(f"Before NMS: {len(detections.xyxy)} detections") detections.xyxy = detections.xyxy[nms_idx] detections.confidence = detections.confidence[nms_idx] detections.class_id = detections.class_id[nms_idx] print(f"After NMS: {len(detections.xyxy)} detections") # Somehow some detections will have class_id=-1, remove them valid_idx = detections.class_id != -1 detections.xyxy = detections.xyxy[valid_idx] detections.confidence = detections.confidence[valid_idx] detections.class_id = detections.class_id[valid_idx] # Segment Anything detections.mask = get_sam_segmentation_from_xyxy( sam_predictor=self.sam_predictor, image=image_rgb, xyxy=detections.xyxy ) # Convert the detection to a dict. Elements are np.ndarray detections_dict = convert_detections_to_dict(detections, classes) # Filter out the objects based on various criteria detections_dict = filter_detections(self.cfg, detections_dict, image_rgb) if len(detections_dict["xyxy"]) < 1: raise SkipImageException("No object detected after filtering.") # Subtract the mask of bounding boxes that are contained by it detections_dict["subtracted_mask"], mask_contained = mask_subtract_contained( detections_dict["xyxy"], detections_dict["mask"], th1=0.05, th2=0.05 ) # Sort the dets by area detections_dict = sort_detections_by_area(detections_dict) # Add RLE to dict detections_dict = post_process_mask(detections_dict) # Convert the detection to a list. Each element is a dict detections_list = convert_detections_to_list(detections_dict, classes) # Skip objects with confidence lower than 0.4 # detections_list = skipbyconfidence(detections_list) detections_list = crop_detections_with_xyxy(self.cfg, image_rgb_pil, detections_list) detections_list = segmentImage(detections_list, image_rgb_pil) detections_list = add_index_to_class(detections_list) if two_class is not None: if len(two_class)==2 and len(detections_list) != 2: raise SkipImageException("Not all objects detected.") if len(two_class)==1 and len(detections_list) != 1: raise SkipImageException("Not all objects detected.") if len(two_class)==3 and len(detections_list) != 3: raise SkipImageException("Not all objects detected.") if len(two_class)==2: detections_two_class = [detections_list[0]['class_name'][:-1], detections_list[1]['class_name'][:-1]] if two_class[0] not in detections_two_class or two_class[1] not in detections_two_class: raise SkipImageException("Not all objects detected.") if len(two_class)==3: detections_two_class = [detections_list[0]['class_name'][:-1], detections_list[1]['class_name'][:-1], detections_list[2]['class_name'][:-1]] if two_class[0] not in detections_two_class or two_class[1] not in detections_two_class or two_class[2] not in detections_two_class: raise SkipImageException("Not all objects detected.") if plot_som: # Visualize with SoM vis_som = draw_som_on_image( detections_dict, image_rgb, label_mode="1", alpha=0.4, anno_mode=["Mask", "Mark", "Box"], ) else: vis_som = None return vis_som, detections_list # Copy the object area from the original image to a transparent background def segmentImage(detections_list, image_rgb_pil): for i in range(len(detections_list)): image_pil = detections_list[i]['image_crop'] mask_pil = Image.fromarray(detections_list[i]['mask_crop']) image_rgba = image_pil.convert("RGBA") transparent_bg = Image.new("RGBA", image_rgba.size, (0, 0, 0, 0)) # Copy the object area from the original image to a transparent background using a mask segmented_image = Image.composite( image_rgba, transparent_bg, mask_pil ) detections_list[i]['image_segment'] = segmented_image return detections_list def skipbyconfidence(detections_list): skip_index = [] for i in range(len(detections_list)): if detections_list[i]['confidence'] < 0.3: skip_index.append(i) for i in skip_index[::-1]: del detections_list[i] return detections_list def add_bbox_and_taggingtext_to_image(image, detections_list): for i in range(len(detections_list)): bbox = detections_list[i]['xyxy'] label = detections_list[i]['class_name'] confidence = detections_list[i]['confidence'] cv2.rectangle(image, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])), (0, 255, 0), 2) cv2.putText(image, f"{label} {confidence:.2f}", (int(bbox[0]), int((bbox[1]+bbox[3])/2)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2) return image def add_index_to_class(detections_list): # If a class appears for the first time, add 0 to the object's class_name, add 1 to the second appearance, and so on class_index = {} for detection in detections_list: class_name = detection['class_name'] if class_name not in class_index: class_index[class_name] = 0 else: class_index[class_name] += 1 detection['class_name'] = f"{class_name}{class_index[class_name]}" return detections_list